Graph-based 3D spatial gene neighborhood networks in gel-embedded single cells

Introduction: Advancements in imaging and experimental technology allow the study of 3D spatially resolved transcriptomics, and improve the representation of the true spatial context of the samples over approximating the samples as 2D. However, introducing the third spatial dimension increases the data complexity. Additional analysis is also needed to further utilize the improved spatial context information. Graph-based deep learning methods have been used to analyze spatial transcriptomics information. For example, graph convolutional network have been used to analyze spatial transcriptomics data in tissue, and identify spatially and transcriptionally unique cell types. However, most ignore subcellular spatial distribution of transcripts. Methods have been developed for inference of cell-cell interaction from transcriptomics data. For example, ligand-receptor pair expression change in cells could be analyzed to infer possible cell-cell interactions. Therefore, we seek to apply graph-based deep learning method to identify potential cell-cell interactions in mesenchymal stem cells cultured in 3D hydrogel.

Materials and

Methods: PEG-MAL macromers were tethered with a 1:4 ratio of macromers and thiolated GFOGER peptide. All components were diluted using PBS++ solution with a pH of 7.4 and 1% HEPES. Covalent conjugation ingredients, a 1:1 mixture of MMP9-degradable VPM peptide to non-degradable DTT molar ratio, were suspended together at a pH of 6.4 HEPES buffer. Cells were suspended in the crosslinker solution. Hydrogels were crosslinked by a mixture of a 1:1 volume ratio of macromers and crosslinker-cell suspension in glass-bottom 96-well plates. The mixture was mixed rapidly via pipet and cured at 37°C for 15 min. Post-curing, the cell media were added to each hydrogel assembly. The media was replenished every three days.
After culturing for four days, the cells were fixed with 4% paraformaldehyde for 10 minutes, and permeabilized in 70% ethanol at -20°C overnight. The mRNA targets were then labeled with the hybridization chain reaction (HCR) protocol and imaged through a fluorescent microscope. Images were analyzed to extract the location of each mRNA molecule. The fluorescent labels were then removed by recombinant DNase I. The next set of mRNA targets were then labeled with the same HCR protocol. (Fig. 1A)

Results, Conclusions, and Discussions: The images were analyzed to detect 3D positions of each transcript. The positions of all detected transcript were combined, and then clustered into spatially distinct subcellular patches. Within each patch, we constructed a nearest-neighbor graph among transcripts in each patch and used permutation analysis to compute the pair-wise proximity score (Fig. 1B, 1C). Both BM-MSCs and UC-MSCs show varying proximity relationships across the same cell (Fig. 1C). To further study RNA subcellular distribution, we established a nearest neighbor graph among all patches in the same cell (Fig. 1C).
We defined each patch as node, and the pairwise gene proximity of the patch as the node features, and connected an edge between each patch and its four surrounding patches to form the nearest neighbor graph. We then applied a graph autoencoder with two convolutional encoders and an inner product decoder (Fig. 1D). The decoder predicts an edge between each pair of nodes based on the output of the convolution layers. The predicted edges were compared with the original graph to train the convolutional layers. The embedded features of each node were shown in color-coded scatter plot with each point represent the position of the node, and the color show the value of embedded feature (Fig. 1E). We then applied a graph neural network (GNN) explainer to identify node and edges with high importance to the embedded features (Fig. 1F).
We then clustered the MSCs based on RNA expression into nine clusters. We define cell-cell contact between cells of the same cluster as homotypic interaction (Fig. 1G), and contact between cells of different clusters as heterotypic (Fig. 1G). We define patches on cell borders that contact a homotypic cells as homotypic patches, and patches that contact a heterotypic cell as heterotypic. We used the GNN explainer to identify important edges and nodes around the border patches, and defined them as cell interaction motifs (Fig. 1H). We then predicted the patch type based on motif embedding and gene proximity. We observed the prediction accuracy of cell interaction motif embedding (Fig. 1I) is worse than that of local gene proximity (Fig. 1J).

Acknowledgements (Optional):